Professional Documents
Culture Documents
Signal Detection Theory (SDT) : Herv e Abdi
Signal Detection Theory (SDT) : Herv e Abdi
Hervé Abdi1
Abstract
Signal Detection Theory (sdt) is used to analyze data coming from
experiments where the task is to categorize ambiguous stimuli which
can be generated either by a known process (called the signal ) or
be obtained by chance (called the noise in the sdt framework). In
particular sdt is used to analyze experiments where a binary answer
(e.g., “Yes” vs “No”) needs to be provided. For example, if we need
to decide if an education program is effective or not, we can use sdt.
Keywords: Signal Detection Theory, False Alarm, Hit, Miss, cor-
rect rejection, d0 , Criterion, Ideal Observer, ROC curve, Type I & II
errors.
1 Overview
Signal Detection Theory (often abridged as sdt) is used to analyze data
coming from experiments where the task is to categorize ambiguous stimuli
which can be generated either by a known process (called the signal ) or be
obtained by chance (called the noise in the sdt framework). For example a
radar operator must decide if what she sees on the radar screen indicates the
presence of a plane (the signal) or the presence of parasites (the noise). This
type of applications was the original framework of sdt (see the founding
1
Cite as: Abdi, H. (in press, 2009). Signal detection theory. In McGaw, B., Peterson,
P.L., Baker, E. (Eds.): Encyclopedia of Education (3rd Ed). New York: Elsevier.
Address correspondence to: Hervé Abdi
Program in Cognition and Neurosciences, MS: Gr.4.1,
The University of Texas at Dallas,
Richardson, TX 75083–0688, USA
E-mail: herve@utdallas.edu http://www.utd.edu/∼herve
1
work of Green & Swets, 1966) But the notion of signal and noise can be
somewhat metaphorical is some experimental contexts. For example, in a
memory recognition experiment, participants have to decide if the stimulus
they currently see was presented before. Here the signal corresponds to
a familiarity feeling generated by a memorized stimulus whereas the noise
corresponds to a familiarity feeling generated by a new stimulus.
The goal of signal detection theory is to estimate two main parameters
from the experimental data. The first parameter, called d0 , indicates the
strength of the signal (relative to the noise). The second parameter called
C (a variant of it is called β), reflects the strategy of response of the par-
ticipant of being more willing to say (e.g., yes rather than no). Sdt is used
in very different domains from psychology (psychophysics, perception, mem-
ory), medical diagnostics (do the symptoms match a known diagnostic or can
they be dismissed are irrelevant), to statistical decision (do the data indicate
that the experiment has an effect or not).
2 The Model
It is easier to introduce the model with an example, so suppose that we have
designed a face memory experiment. In the first part of the experiment, a
participant was asked to memorize a list of faces. Then the participant is
presented with a set of faces one at a time. Some of these faces were seen
before (these are old faces) and some were not seen before (these are new
faces). The task is to decide for each face if this face was seen (response Yes)
or not (response No) in the first part of the experiment.
What are the different types of responses? A Yes response given to an
old stimulus is a correct response, it is called a Hit; but a Yes response to
a new stimulus is a mistake, it is called a False Alarm (abbreviated as FA).
A No response given to a new stimulus is a correct response, it is called a
Correct Rejection; but a No response to an old stimulus is a mistake, it is
called a Miss (abbreviated as FA). These four types of response (and their
frequency) can be organized as shown in Table 1.
The relative frequency of these four types of response are not all inde-
pendent. For example when the signal is present (first row of Table 1) the
proportion of Hits and the proportion of Misses add up to one (because when
the signal is present the subject can say either Yes or No). Likewise when the
signal is absent, he proportion of FA and the proportion of Correct Rejection
2
Table 1: The four possible types of response in sdt
reality Yes No
3
larger than a predefined threshold.
Sdt also assumes that the stimuli generated by the noise condition vary
naturally for that hidden variable. As is often the case elsewhere, sdt, in
addition, assumes that the hidden variable values for the noise follow a normal
distribution. Recall at this point, that when a variable x follows a Gaussian
(a.k.a Normal) distribution, this distribution depends upon two parameters:
the mean (denoted µ) and the variance (denoted σ 2 ). It is defined as:
½ ¾
1 (x − µ)2
G(x, µ, σ) = √ exp − . (1)
σ 2π 2σ 2
In general within the sdt framework the values of µ and σ are arbitrary and
therefore we choose the simpler values of µ = 0 and σ = 1 (other values will
give the same results but with more cumbersome procedures). In this case,
Equation 1 reduces to
½ ¾
1 1 2
N (x) = √ exp − x . (2)
2π 2
Finally, sdt assumes that the signal is added to the noise. In other words,
the distribution of the values generated by the signal condition has the same
shape (and therefore the same variance) as the noise distribution.
Figure 2 illustrates the sdt model. The x-axis shows the intensity of un-
derlying hidden variable (e.g., familiarity for the face example). As indicated
above, the distribution of the noise is centered at zero (i.e., mean of the noise
is equal to zero), with a standard deviation of 1. So, the standard deviation
of the noise is equivalent to the unit of measurement of x. The distribution
of the signal is identical to the noise distribution, but it is moved to the right
of the noise distribution. The distance between the signal and the noise dis-
tributions corresponds to the effect of the signal (this is the quantity that is
added to the noise distribution in order to get the signal distribution): this
distance is called d0 . Because the mean of the noise distribution is zero, d0 is
equal to the mean of the signal distribution.
The strategy of the participant is expressed via the choice of the threshold.
There are several ways of expressing the position of this threshold, among
the possible candidates we will mention four of them denoted B, D, C and β.
The first quantity B (sometimes called ϑ) gives the position of the threshold
on the x-axis. In the example illustrated in Figure 2, this value is equal to
2, and so the participant corresponding to this figure has decided that any
4
Figure 1: The model of sdt.
stimulus with a value of x larger than 2 comes from the signal distribution
and is given the response Yes. The position of the threshold can also be given
relative to the signal distribution (because the noise has zero mean, B is the
distance of the threshold relative to the noise distribution), as the mean of
the signal is equal to d0 we can compute D as D = d0 − B (a value equal to
−1 in our example).
The most popular way of expressing the location of the threshold, how-
ever, is neither from the distribution of the noise nor the distribution of the
signal but relative to what is called the ideal observer. The ideal observer
minimizes conjointly the probability of a Miss and of an FA. When each type
of errors has the same cost, the criterion of the ideal observer is positioned
on the average of the means of the signal and the noise distribution. In our
example, the threshold of the ideal observer would be equal to 12 d0 = 12 = .5.
The value of C is the distance from the actual threshold to the ideal ob-
0
server, it can computed as C = B − d2 = 2 − .5 = 1.5. The sign of C reveals
the participant’s strategy: when C = 0, we have the ideal observer; when
C is negative the participant is libéral (i.e., responds Yes more often than
5
Table 2: The probability or the four possible types of response according to
to Figure 2.
6
(i.e., area under the normal distribution) of finding a value larger than 2
with a normal distribution of mean 0 and variance 1 (this can be computed
with most statistical packages or from Tables such as the ones given in Abdi,
Edelman, Valentin & Dowling, 2009). This quantity is also called the prob-
ability associated to the value 2, in our example it is equal to .0228. Along
the same lines, the probability of a Hit is obtained as the probability (i.e.,
area under the normal distribution) of finding a value larger than 2 with a
normal distribution of mean 1 (i.e., the mean of the signal) and variance 1,
this is equivalent of finding the probability (i.e., area under the normal dis-
tribution) of finding a value larger than 2 − 1 = 1 with a normal distribution
of mean 1 − 1 = 0 and variance 1. This value is equal to .1587.
4 SDT in practice
The previous example was describing the performance of a participant who
behaved according to the sdt model. However, in practice we do not known
the values of the parameters of sdt, but we want to estimate them from
the performance of the participants. In an experimental paradigm the only
observable quantities are the participant’s responses from which we can derive
the number of hits and FA’s.
To illustrate this problem suppose that we want to evaluate the perfor-
mance of a wine taster whose task is to detect if a wine labelled as made from
“Pinot Noir” has been tempered by the addition of some Gamay (generally
considered an inferior grape). Here, the signal corresponds to presence of
Gamay. Our wine taster tasted (blindfolded) twenty glasses of Pinot, (half
of them tempered with some Gamay and half without). The results are
reported in Table 3, and show that the proportion of Hits and FAs are re-
spectively .9 and .2. In order to find the values of d0 and the criterion, we
need to inverse the formulas given above (i.e., Equation 3–5). We need one
new notation: for a normal distribution with zero mean, we denote by ZP
the value of the normal distribution whose associated probability is equal to
P (e.g., Z.025 = 1.96). We denote as ZH et ZF A the values corresponding to
the proportions of Hits and FAs. With these new notations and after some
(minor) algebraic manipulations we find the following set of formulas. The
estimation of d0 is obtained as
7
Table 3: The performance of a wine taster trying to identify Gamay in a
Pinot Noir wine.
8
6 Bibliography
The classic work on sdt is Green and Swets (1966), a basic introduction
is McNicol (1972), two recent comprehensive references are Macmillan and
Creelman (2005) and Wickens (2002).
References
[1] Abdi, H., Edelman, B., Valentin, D., & Dowling, W.J. (2009). Experimen-
tal design and analysis for Psychology. Oxford: Oxford University Press.
[2] DeCarlo,L.T (2005). A model of rater behavior in essay grading based on
signal detection theory, Journal of Educational Measurement, 42, 53-76.
[3] DeCarlo, L. T., & Luthar, S. S. (2000). Analysis and class validation of
a measure of parental values perceived by early adolescents: An applica-
tion of a latent class model for rankings. Educational and Psychological
Measurement, 60, 578–591.
[4] Green D.M., Swets, J.A. (1966). Signal detection theory annd psy-
chophysics. New York Wiley.
[5] McDermott, P.A., Fantuzzo, J.W., Glutting, J.J., Watkins, M.W., & Bag-
galey, R.A. (1992). Illusions of ipsative assessment of children’s ability.
Journal of Special Education, 25, 504-526.
[6] Quantifying the information value of clinical assessment with signal deci-
sion theory, Annual REview of Psychology, 50, 215–241.x
[7] Macmillan, N.A., Creelman, C.D. (2005). Detection theory: A user’s guide
(2nd edition). Mahwah (NJ): Erlbaum.
[8] McNicol, D. (1972). A primer of signal detection theory. London: George
Allen & Unwin.
[9] Wickens, T.D. (2002). Elementary signal detection theory. Oxford: Oxford
University Press.
Author’s biography
Hervé Abdi received an M.S. in Psychology from the University of Fran-
che-Comté (France), an M.S. (D.E.A.) in Economics from the University of
Clermond-Ferrand, an M.S. (D.E.A.) in Neurology from the University Louis
Pasteur in Strasbourg (France), and a Ph.D. in Mathematical Psychology
9
Figure 2: Hervé Abdi
10